Transformer-Based Detection of Autism Spectrum Disorder using fMRI Data

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by difficulties in social interaction, communication, and repetitive behaviors. Its diagnosis relies on behavioral assessments, which can be time-consuming and imprecise due to their subjective nature and the wide variability of symptoms among individuals. When not identified in early life, diagnosing ASD becomes even more challenging, as people may develop mechanisms that mask their symptoms. Thanks to advances in neuroimaging techniques, such as Functional Magnetic Resonance Imaging (fMRI), it has been possible to obtain Blood Oxygen Level Dependent (BOLD) signals that reflect brain activity during rest or task performance. Machine Learning and Deep Learning methods have been applied to analyze these complex datasets, aiming to identify patterns associated with ASD and improve diagnostic accuracy. In previous works, the Long Short-Term Memory (LSTM) network have shown the most promising results in classifying ASD from fMRI data by studying temporal dependencies in the BOLD signals. What we wanted to investigate in this thesis is the potential of Transformer-based architectures, trying to understand if they can outperform the LSTM model. Transformer models have the capability to capture long-range dependencies in sequential data, making them suitable for analyzing the temporal dynamics of fMRI signals. Resting-State fMRI (rs-fMRI) data are taken from the Autism Brain Imaging Data Exchange (ABIDE) I, a large-scale multi-site dataset which includes BOLD time series data from both ASD-diagnosed and typically developing (TD) individuals. A Simple Transformer Encoder (STE) and a Long Short-Term Memory (LSTM) model are proposed to establish a baseline. BOLD Transformer (BolT) which integrates a windowed mechanism within the transformer architecture in order to better capture local temporal patterns in the fMRI time series, has been explored, obtaining higher performance compared to the STE. From the results obtained with the study of these two models is emerged that BolT almost reached the performance of the LSTM, but did not surpass it, suggesting that further refinements to the Transformer architecture may be necessary to exploit them and try to overcome the LSTM in this specific application or that LSTM is still a strong basline. To interpret the decisions made by the Transformer-based models, Layer-wise Relevance Propagation (LRP) can be used to compute relevance scores for each input feature. The application of LRP aims to provide the identification of the most relevant brain regions contributing to the classification of ASD. The interpretability method proposed by BolT has been adapted to our work in order to make it fairer and it did not show consistent results, while relevance scores obtained by a fixed LRP method for Transformer architecture have been more stable when applied to BolT, leading to better results. LRP is fixed to work with Transformer architectures because it has demonstrated that several parts of the model can break the conservation property of LRP if not treated properly. To further investigate the importance of these identified features and to validate whether features with lower relevance scores can be discarded to enhance model performance and mitigate the risk of overfitting, we implemented and applied several input masking techniques. Specifically, two types of input masking were considered based on the relevance scores obtained from the fixed LRP: soft masking, where less relevant features are down-weighted, and binary masking, where only the most relevant features are kept given a certain threshold. An extension of the binary masking has been implemented, completely removing the less relevant ROIs from the input data. These masking techniques did not show significant improvements in model performance, indicating that even masking or removing a small set of features may lead to loss of important information.